ResearchBenchmarksSafety Representation Extraction on Conversation level datasetSafety Representation Extraction on Conversation level datasetFollow100Accuracy (MCA)ReGA96.8897.6998.599.31Jun 2, 2025Evaluation resultsMethodMethodLinksAccuracy (MCA)Accuracy (MFP)AUROCReGAModel=llama-13BModel=llama-13B2025.0610010099.8ReGAModel=qwen-14BModel=qwen-14B2025.069910099ReGAModel=vicuna-13BModel=vicuna-13B2025.06979499